Video transcript
Transcript from available captions; it may contain errors.
Claude built an entire operating system in 8 days. And I think people are missing the biggest part of this story. Anthropic recently released Claude Fable 5.2 and Mythos 5.2. And instead of having an entire development team write the code, it had AI agents working together. But the operating system wasn't even the only surprise. Anthropic also gave these models problems that normally require experienced engineers and scientists. And some of the results are genuinely difficult to believe. And while everyone is talking about Claude, Google has finally released its most powerful AI model, Gemini 4 Carbon. This model can handle complex software projects, check its own mistakes, and keep iterating until things work. If these capabilities hold up, companies could build software much faster. But OpenAI has made a very unusual decision this week. It developed a cyber security model with capabilities so sensitive that it's restricting who can use it and the reason behind that decision is worth understanding. [music] At the same time, OpenAI is bringing businesses directly into Chat GPT and Elon Musk's Grock can now handle those frustrating customer support calls on your behalf. These are only a few of the 18 biggest AI updates this week. At the end, we've put Claude artifacts to the test, and we'll show you what they can do, how well they work, and how you can use them yourself. As always, I've put all the important prompts, tools, and links inside our free staying ahead WhatsApp community. The link is in the description. Now, let's get into the video. Anthropic just launched Claude Fable 5.2 and Mythos 5.2. These models are built for some of the most difficult coding, scientific research, and engineering tasks. And the results Anthropic has shared are pretty impressive. In one experiment, an engineer gave Fable 5.2 an empty folder and asked it to build an entire operating system from scratch. Fable brought together more than 1,000 AI agents and kept them working for 8 days. By the end, they had built an operating system that could actually run on a laptop, complete with Windows, file storage, and a login screen. Ananthropic took things beyond software. Using simple text descriptions, Fable designed three electronic devices, a voice assistant, a USB speaker phone, and a virtual pet. It designed everything from the circuit boards and outer cases to the software running inside them. A hardware company then assembled these devices and all three worked. Anthropic also tested Fable with real robots. The robots managed to solve a puzzle cube, unscrew bottle caps, and even drop pictures despite not being specifically trained for those tasks. And its performance on a robotics benchmark more than doubled compared to the previous version. What's even more interesting is how Fable performed in scientific research. Anthropic used another 1,000 Fable agents to convert the famous proof of Fermat's last theorem into a format that computers could verify. They completed the entire process in just 30 hours with very little human involvement. Fable also helped researchers tackled two previously unsolved problems in black hole physics. One of them involved completing a calculation that scientists had proposed more than 20 years ago back in 2003. And mythos showed some interesting results in medical research. It helped predict the complete structures of tiny virus shells that scientists study for gene therapy. According to anthropic, mythos accurately predicted 28 out of 34 structures in its tests. Now coming back to coding, which is where most of us might actually use Fable, a team of Fable agents converted one of Anthropic software services from Python to Rust, making it start 26 times faster. Anthropic says Claude now writes more than 90% of the code changes added to its own core software, although humans still guide and review the work. Mythos has also shown strong results in cyber security. In one internal test, it successfully identified 89.6% of known security weaknesses by examining older versions of software. And on the Frontier Code 2.0 coding benchmark, Fable 5.2 2 scored 49.1% at its highest effort setting, outperforming Claude Opus 5.5 and GPT6 Astra in anthropics comparison. Now, there's an important difference between Fable and Mythos. Both share the same underlying AI, but Fable comes with stronger safety restrictions and is designed for general use. Mythos, meanwhile, is available only to approved organizations working on sensitive cyber security and medical research. Fable 5.2 2 is already available on paid cloud plans and Anthropic has also made it 20% cheaper through the API with up to 30% faster output generation compared to its predecessor. Google's most awaited AI model has finally launched and it's called Gemini 4 Carbon. This is Google's strongest model yet. Built for difficult coding tasks, deep research, cyber security, and AI agents that can work on complex projects. Carbon can check its own code and screen while building applications or even 3D worlds, identify mistakes and fix them along the way. Google has also introduced/swarm in anti-gravity where multiple AI agents can work together while you track their progress. Carbon is now available through Google AI Ultra, Anti-gravity, Google AI Studio, and the API with introductory pricing matching Gemini 4 Argon. And speaking of Google, we've also recorded a podcast with Logan. We had a really interesting conversation about AI and what's coming next. That episode is dropping soon, so stay tuned. OpenAI just launched Chat GPT images 2.6 and you can now create animated gifts from simple prompts. Look at these. You can create a dancing cat, a disco frog, or even a dog doing cartwheels. And once you've created an animation, you can keep making changes to it. For example, you can add pink lasers to the bird's sunglasses. And Chat GPT updates the animation without starting from scratch. You can even change the animation speed. And look at these animals performing like K-pop stars. All created from simple prompts. But GIFs aren't the only improvement. OpenAI says image generation is now 50% faster, and characters stay more consistent when you edit them. You can also save reference images of people, pets, or characters and reuse them in future chats without uploading them again. Images 2.6 is now available globally across all chat GPT plans. OpenAI just launched GPT6 Cyber, a new AI model built specifically to find security weaknesses in software. It can go through code, find problems that hackers could use to break into systems, and help developers fix them. And the results are pretty impressive. OpenAI says the model developed 948 working attacks against security weaknesses in Linux and discovered 81 serious security problems in the Firefox browser that were previously unknown. Mosilla has already released fixes for many of them. But OpenAI is also making sure the model stays within its limits. In one test, GPT6 Cyber only checked systems it had permission to test 99.93% of the time. For now, it's available only to approved security researchers through OpenAI's Daybreak Red program. OpenAI is also expanding Daybreak Blue to help more security teams find and fix problems. Grockbot can now make phone calls on your behalf. For example, it notices your internet plan is renewing in 3 days and offers to cancel it. You agree, but tell Drockbot not to accept any discounts. Before dialing, it asks for your approval. Once connected, it speaks directly to customer support and requests the cancellation. The support agent offers 10% off for a year, but Grockbot sticks to your instructions. It mentions that you found a better deal with Starlink, rejects the discount, and gets the plan cancelled. You can follow the live conversation and even take over the call whenever you want. Once the call ends, Grockbot sends a confirmation directly to your chat. So, Grockbot handles the entire conversation with customer support while you stay in control. Anthropic has launched three new features that turn Claude into a workspace for your entire team. First, there's Claude whiteboards. If you're planning a product launch, you can describe your idea and Claude creates a visual timeline with tasks, milestones, and illustrations that your team can edit together. Next is Claude Sheets, where you can create spreadsheets, build sales forecasts, and work with formulas. For example, Claude can use last year's sales data to predict how many orders you might receive this year. Then there's Claude data, previously called dashboards. It can turn your numbers into interactive dashboards, charts, and even visual stories that you can share with your team. Anthropic has also upgraded its connectors. So tools like Adobe, Atlassian, Box, and Notion can now open directly inside Claude. So Claude is becoming a place where you can plan, build, analyze, and collaborate without constantly switching between different apps. These artifacts are so amazing, we decided to test them ourselves. We've added a step-by-step tutorial at the end of this video, so you can try these features yourself. But first, let's get into the other big AI updates of the week. OpenAI has launched businesses in Chat GPT, and this could change how customers discover your business online. If you run a toy store, you can now claim your business profile inside ChatGpt, update your contact details, and upload your products along with their prices. You can now also create your own business agent. So, when a customer asks Chad GPT for a drawing gift under $30, your agent can recommend products from your store and answer questions about them. You also get a dashboard showing how often your business appears in ChatGpt, which topics bring customers to you, and how many people visit your website. So, Chat GPT is becoming another place where customers can discover businesses and interact with them directly. Perplexity just launched teammates, AI agents. You can give ongoing responsibilities at work. For example, if you run an online store, you can create a store manager and connect it to Shopify and Gmail. You can put it in charge of inventory and it checks stock, reviews orders, and prepares reorders for your approval. These teammates [music] can even work together. An engineering teammate can fix a password reset problem and then pass it to another teammate to test whether the fix works. They can also use your existing projects, tools, and memory. So you don't have to explain everything from scratch. You can have different teammates handling ads, inventory, customer support, and coding, all working in the background. And when you give them feedback, they remember it and apply it to future tasks. Your co-workers can even message them directly in Slack, where a legal teammate can review contracts and prepare suggested changes. So instead of giving AI a new prompt for every task, you can now give it an entire responsibility and let it keep the work moving. Meta just launched Muse Ember, its smallest and cheapest AI model yet. It's designed to handle the smaller tasks that AI agents perform hundreds of times, like reading documents, making tool calls, extracting information, and fixing code. and it can understand text, images, audio, and video with a massive 1 million token context window. News Ember costs just 10 cents per million input tokens and 50 cents per million output tokens for standard requests. Meta claims it delivers better intelligence for every dollar than Claude Haiku 5.5 and GPT6 Luna based on artificial analysis benchmark comparisons. And for bigger projects, you can use Muse Wonder for planning, while Ember handles the smaller tasks in the background. Google just launched Gemini Omni 1.2 Flash, and it can now generate videos up to 15 seconds long. For example, you can upload a single image of a honeycomb, and Gemini turns it into a fourshot science explainer. It shows the queen bee, the different stages of a bee's life, and how nectar turns into honey, complete with labels that follow the action. The update also keeps characters and scenes more consistent across different shots. You can make changes using follow-up prompts, and Google has improved speech and lip-sync to sound more natural. Google has also reduced API pricing by 30%. Bringing it down to just 7 cents per second for 720p video. Anthropic just published new research on how Claude can remember information more effectively across different tasks. One of the ideas is called dreaming. It allows Claude to review its previous work, correct mistakes in its memory, and organize important information for future tasks. Anthropic found that remembering everything isn't always better. In one Pokémon test, an older Clawude model saved 413 memory files, but earned only one badge. A newer model saved just 10 carefully selected files and earned four badges. Anthropic also found that Claude could build its own wiki from thousands of documents, making important information easier to find later. So, the goal isn't to make Claude remember every little detail. It's to help Claude remember what matters, learn from past mistakes, and use that knowledge in future tasks. Anthropic just launched personal finance in Claude. It can now check your spending, show where your money goes, and help you plan your budget. If you're planning a family vacation to Rome, you can simply connect the bank accounts you want to share through Plaid. Then you can ask something like, "What can we even afford?" Claude goes through your transactions, breaks down your spending into simple charts, and shows how much goes toward fixed bills, and where you can actually cut back. Claude also notices your car loan is almost paid off, which means you'll soon have an extra $348 every month. It even spots $180 in credit card interest and fees. Now, you can ask if that extra money could extend your Rome trip from 7 days to 10. Claude works out a savings plan and maps everything onto a calendar so you can see how the numbers add up. And once you've figured out your budget, you can even ask Claude to find hotels that fit within it. It pulls up two options with their nightly prices so you can choose where to stay. And importantly, Claude can only read the accounts you connect. It cannot move or invest your money. OpenAI just announced plans to acquire Pierre, a company that builds tools for software developers. And if you've used Codeex to browse files or review code changes, you've already used Pierre's technology. Its tools help you navigate large coding projects and see exactly what has changed. Pierre also built code storage, which lets apps store code and track its history. OpenAI now wants to bring more of these tools into Codeex so developers and AI agents can build, review, and manage code in the same place. Pierre's open- source tools will remain open source, and existing code storage customers can continue using the service after the deal closes. So, OpenAI is slowly turning Codeex into a complete workspace for software development. Meta just launched spaces in Muse. You can now create your own planners, trackers, and dashboards simply by describing what you want. For example, you can create a family dashboard that brings together your calendar, school pickups, and lunch plans. And if you want to add a section for movie night, Muse updates the dashboard through the same chat. There's also an example called Jolly Mall. It takes posts you've saved on Instagram and turns them into shopping recommendations displayed like a shopping catalog from the '90s. You can even browse individual products in that same retro style, and you can customize the entire app around your preferences. You can move your favorite spaces into the main tabs and rearrange everything based on what you use most. For example, you can type weekend mode and Muse brings your family dashboard and trip planner to the front while moving workrelated tools out of the way. These spaces can also keep updating through your connected apps. So instead of adjusting your routine around the app, Muse now lets you build and organize the app around your own routine. Anthropic just published new research on how AI could help scientists understand how the world works. For example, scientists use mathematical equations to understand how brain cells send signals. AI agents can now help build these equations, test them against real data, and improve them when the results don't match. They can even explore different explanations for the same scientific problem. And that's important because two different explanations can produce similar results even if scientists don't fully understand why. By comparing them, AI could help researchers identify what they still don't know and which experiments to run next. For now, the technology is still early, but it could eventually help scientists study much more complicated systems, including entire brains. Google just upgraded flashcards in Gemini Notebook with three new features. You can now add images to your flashcards. For example, if you're studying art, you can see the painting, answer a question, and flip the card to check. You can also turn on a stopwatch to test how quickly you remember each answer. And on the mobile app, you can set reminders so you don't forget your next study session. So, you can revise with images, track your speed, and keep up with your studies. Notion just launched shared memories for AI agents. This means chat GPT Claude and other AI tools can now access the same information about your work. For example, you tell Claude code to move your website launch to next Friday because the mobile version needs more testing. Claude saves that decision directly in notion. Now when you open chat GPT and ask it to write a project update for your team, you don't need to explain everything again. Chat GPT retrieves the information saved by Claude and drafts the update with the correct launch date and the tasks your team still needs to finish. So your AI tools can now share what they've learned even when you're switching between them. India's server AI just launched Prism, a new system one model built on its 105B model. Unlike a regular chatbot, Prism is designed to make quick decisions instead of writing long answers. For example, it can decide whether a customer complaint should go to billing, delivery, or technical support. And its speed is impressive. In Serv's tests, Prism scored 71.8 across three benchmarks with a median response time of just 44 milliseconds. That made it the highest scoring model under 100 milliseconds among the 25 models tested. It also scored 82.5 on Indian language tests and tells you how confident it is about each decision. Even changing the order of the options doesn't change its answer. Serv has released Prism with open weights on hugging face so companies can download, customize, and run it themselves. There are three new things in Claude that you should know about and I am going to walk you through each one. Number one is sheets, a spreadsheet that Claude builds and then works in with you. Number two is data which turns your numbers into a dashboard that shows where every figure came from. And number three is whiteboard, a board that you and Claude both draw on. Let us start with number one where I gave sheets one raw file and one line of instructions. Open the Claude desktop app and click artifacts in the sidebar. You get six things Claude can make for you here which are docs, slides, sheets, design, data and whiteboard. And two of them, sheets and whiteboard, carry a beta tag. Hover over sheets and Claude describes it as a grid with formulas and formatting that you and Claude fill in together. So, how do you get your own data in? Because there are two ways. And the first one is to doubleclick the sheet style, which opens a blank spreadsheet. Along the bottom, you can import a CSV file, an Excel file, or a Google sheet. Or you can ask Claude. Pick your CSV. And before anything lands, Claude reads the file and tells you what is in it, which for mine was one country per row, 126 rows, and seven columns. Hit apply, and the whole file is sitting in the sheet. From here, you could click ask Claude and describe a change. But there is a faster way to do the whole thing in one go. Go back to the sheet style and choose start with a prompt. Drop in your file and mine is real public data from our world in data on how many hours people work and how much they earn in 126 countries. Then tell it what you want and I asked it to sort by annual hours with the highest first. Add a column that ranks each country and highlight India, the United States and Germany. This ran on Opus 5.5 at medium effort. So nothing special on the model side. The sheet opens next to the chat and that little orange tag moving around the grid is Claude's own cursor. It writes the headers first and puts the new rank column right at the front. Then every country lands in order with the longest working hours at the top. And then it goes back and highlights the three countries I asked for. Start to finish. That took a little over 2 minutes. [snorts] So what does the data say? Because India comes in 10th out of 126. The United States is 84th and Germany is right at the bottom. Germany works the fewest hours of any country on this list. That means India works about 1,50 more hours a year than Germany. It also told me two things I never asked about. That India is tied with Bangladesh, so they share rank 10. And that the rank is a formula, so it recalculates if a number changes. Is this an actual spreadsheet or only a table that looks like one? Well, type an equal sign in any cell and the formulas come up the way they do in Excel. The toolbar has charts and conditional formatting. Rightclick any cell and you can comment, sort, filter or freeze. And a comment is how you hand that one cell to Claude. When you are done, you can export it as an Excel file or a CSV or send an editable copy straight to Google Sheets. It is private to you by default, so nobody sees it until you share it. To share it, you invite someone by email, choose what they can do, add a note, and send. If you already have a messy file and you know what you want from it, this is the quickest way I have found to get a clean sheet without touching a formula. But a sorted list still does not answer the real question, which is whether the countries that work more actually earn more. And that is a job for number two, data. Open the data tile. And before it builds anything, it asks you one question. Where do your numbers come from? You can import a file or you can use a connector. And Claude's own label on the connector option is that the numbers stay upto-date. The connector list covers the places a business already keeps its numbers like HubSpot, Salesforce, and Stripe. Then you choose what you want back and there are four formats which are a dashboard, a visual analysis, a scrollty teller, and a visual short. You can also pick a design system, which is your colors, your fonts, and your components saved once, so everything Claude makes looks like your brand. I already have one called Attellier, so I picked that and went with files this time. You attach your files right in the message box, and I gave it the same country file as before, plus a second one with the hours every country has worked each year since 1970. Then you ask it one clear question and mine was, do countries that work more hours earn more? I wanted three things on it. A chart of hours against income, the 10 longest and shortest working countries, and how hours have changed over time in five countries. Claude tells you upfront that this one usually takes a few minutes. A little over 5 minutes later, the page appears and it answers my question in its very first word, which is no. Across 126 countries, the longer people work, the lower their income tends to be. The quarter of countries that work the longest earn about $12,000 a person and the quarter that work the least earn about $47,000 which is four times more. It also found the exceptions on its own because only Singapore, Qatar, the UAE, Hong Kong, and Kuwait manage both long hours and high income. The tables match what the sheet told us with Germany working the least and it caught the same tie between India and Bangladesh and over time South Korea cut about a thousand hours a year since 1970 while India's hours barely moved. Can you trust those numbers? Well, every one of them has a source you can open all the way down to the formula. If you want it in another format, you pick one from the design panel and I chose the scrollty teller. Under 4 minutes later, it came back as a draft, which means the live dashboard had not changed for anyone else. The same findings now read like a story you scroll through, one step at a time. I like the dashboard more, so I discarded the draft and the original was still there untouched. From the export tab, you can also connect it to tools like Amplitude, Mix Panel, Data Dog, and Graphfana. for a file you already have. 5 minutes gets you a dashboard that states its answer and shows where every number came from. Number three is the one for thinking out loud, which is whiteboard. Open it and you get an empty board with the tools along the bottom. And if you've used Fig Jam or Miro, it will feel familiar. You can drop sticky notes and keep adding more in any direction with one click. You can draw with the pen or you can type what you want and let Claude draw it, which is how I got a circle. Can Claude read what is on the board? Well, you can comment on anything and tick send to Claude. I wrote, "This is so cool." And Claude's reply was honest because it said the comment was only praise, so it left the board alone. Now, for a real job, you can drop a file straight onto the board. And I used one on private AI investment from 2013 to 2025. I asked for one chart, three big numbers, and a card each for the United States, China, and Europe. About a minute and a half later, this is the board. Its headline is that the United States took 83 cents of every private AI dollar in 2025. That is about 240 billion out of 290 billion worldwide. Europe is at a record and has been ahead of China since 2023, while China sits 60% below its 2021 peak. And everything on it is a real object. So, I can select the 83%, move the chart, or rewrite a card by hand. Claude says it checked every figure against the file, and I checked them, too, and they match. A whiteboard is really for planning, though. So, with no file at all, I asked for a product pipeline for a small notes app. It drew eight steps from collecting ideas to measuring with an arrow looping back that says, "What you learn becomes the next idea." It also told me it did not know my tools, so it picked its own examples. And I can edit any of them right on the board. When it is ready, you can share the link or export the whole board as an image. So that's all three Claude Artifacts features we tested. We started with raw data, built a working spreadsheet and a dashboard, and even used whiteboard to plan an entire project. We've also checked the numbers, tested the editing features, and seen what you can actually do with the results. One thing to remember is that Sheets and Whiteboard are still in beta and all three features require a paid Claude plan. So after trying Claude Artifacts, here's what stands out. You can plan projects, build spreadsheets, [music] and create dashboards all inside Claude without constantly switching between different tools. And if you want to try these features yourself, I've shared all the prompts, the step-by-step guide, and the links inside our free staying ahead WhatsApp community. The link is in the description. And if you want to see how AI agents are actually being used inside a billiondoll company, watch our conversation with Notion's co-founder, Akshai Kotari. He showed us how Notion uses Claude, Codeex, AI agents, and a second brain to get work done. I definitely recommend watching that next. And we have another exciting podcast coming up with Logan from Google. That conversation is dropping soon, so stay tuned. And finally, 70% of the people watching this channel are not subscribed yet. So if you want to catch our next AI updates video and the upcoming podcast with Logan, make sure you hit that subscribe button. I'll see you in the next one.
Assuntos: Agentes de IA, Assistente pessoal, Muse. Publicado no YouTube em 2026-11-11.
Comentários
Postar um comentário